Image recognition device and image recognition program

Abstract

An image recognition device adds a margin region filled with predetermined image data to a periphery of a captured image captured by a camera to create an expanded image larger than the captured image. When a person is too close to the camera, a part of a person image protrudes from the captured image, but a large detection window which also includes the margin region is set, whereby a window image including a protruding region is taken out of the expanded image. The window image lacks a protruding part of the person image, but it is an image showing an entire body of the person image. The image recognition device stores many reference images assuming various states of a person. The image recognition device extracts features from the window image, and compares them with a feature of the reference image, thereby recognizing the person image.

Description

TECHNICAL FIELD[0001]The present invention relates to an image recognition device and an image recognition program and, for example, to those which recognize an object shown in an image.BACKGROUND ART[0002]In recent years, image recognition technologies have been used in various fields of industries or societies, for example, to recognize an object from a captured image for control when a camera is mounted in a robot or a vehicle or to recognize images of people and thereby observe a flow of the people when a camera is installed a street.[0003]Some of these image recognition technologies are carried out by storing a reference image showing features of an object or a non-object (required for recognizing that an image is not the object) in an image recognition device in advance and comparing features of the reference image with those of a camera image.[0004]As such image recognition technologies, there is an “object detection device” of Patent Literature 1.[0005]This technology is con...

AI Technical Summary

Benefits of technology

The invention allows for detecting a target even if it sticks out of the image. This is achieved by expanding the detection window region beyond the image.

Problems solved by technology

Further, when a technique using a luminance gradient such as HOG as features of an image is adopted, since an edge portion 100 where the head region is severed is an end portion of the image, a calculation of the luminance gradient, i.e., extraction of the features is difficult, and any contrivance for the calculation must be applied.

As regards such a problem, a method for creating a reference image in which a part of a person protrudes to the outside of an image and storing it in the image recognition device can be considered, but the reference images in which a part of the person protrudes to the outside of each image in various conformations must be stored to carry out this method, an amount of data to be stored becomes enormous, and implementation of the method is difficult.

Specifically, there are a plurality of (approximately thousands of) reference images (images each showing a whole) to one object, and various kinds of images with missing upper and lower sides, various kinds of images which lack a right side, various kinds of images which lack a left side, and various kinds of images which lack a vertical or lateral side must be prepared and stored for each image, thereby resulting in an enormous amount.

As described above, the conventional technologies have a problem that a recognition target cannot be recognized when a part thereof protrudes to the outside of an image.

In this manner, even if a demand for recognition at a short distance is high, the recognition is impossible since an entire recognition target does not fit in an angle of view of a camera.

Images